Council position · Sovereign AI

When a foreign order can switch off your AI, what do you actually own?

On the evening of June 12, 2026, the United States ordered a major AI company to cut off its two most capable models to every foreign national. Because the company could not separate foreign users from the rest in real time, it shut the models off for everyone, worldwide, the same night. People in Canada who were using the tool an hour earlier simply lost it. For First Nations, the lesson is not which model to use. It is that the model is the one layer you do not control, and sovereignty lives in the layers you do.

From Russell Jones, a Sagamok member, for OIATC. This is OIATC's position, and general information, not legal advice. June 13, 2026.

What happened

The US government sent the order under export-control authority. It applied to any foreign national, whether inside or outside the United States, including the company's own foreign-born staff. The company complied while disputing the reasoning, said it believed the concern was a narrow technical one, and said it expects to restore access. Every other model from that company kept working.

Treat the specific dispute as unsettled. That is not the point. The point is what the event revealed about the shape of the thing.

Why this is not a one-off

The most capable AI models are built by a small number of companies in one or two countries, and they run as a service you reach over the internet. That arrangement is convenient, and most of the time it works. But it means your access is a permission, and a permission can be changed by a government you did not elect, in a country you do not live in, overnight, with no notice to you. This is the same lesson as where your community's data actually lives, one layer up. When the important parts sit outside the community, the community cannot answer the basic question of whether they will be there tomorrow.

The three layers, and which ones you can own

An AI system that serves a community is not one thing. It is three layers stacked on top of each other, and they are not equally in your hands.

Put plainly: the model is the swappable layer. Lose one, you can run another. What must not be swappable from under you is your community's data, and the ground it runs on.

The solution: own the layers you can

If the model is the swappable layer, sovereignty does not live in the model at all. It lives in the two layers underneath, the ones a community can actually hold. The fix is to build so the part that can be revoked is the part you can replace, and to own everything beneath it. Three parts:

• Governed data. A consent-gated corpus held under OCAP and CARE, so no community data trains anyone's model without consent and a share of the benefit.
• Open, swappable models you run yourself. Open-weight models a community can download and run on its own ground, so a model switched off from outside is swapped for another without losing anything that matters.
• Indigenous-controlled hosting. The system runs on infrastructure the community governs, with data residency in the community or in Canada, not on ground someone else can revoke.

Delivered by the sovereign stack OIATC builds and runs: Waaseyaa, the framework underneath, and Anokii, the working reference of this pattern in use today.

What OIATC will not build, and why

OIATC will not build a frontier model. A single training run for one of these systems costs in the hundreds of millions to over a billion dollars, and that is a race for a handful of large companies and governments. Chasing it would be the wrong use of a community's resources, and we will not pretend otherwise. That is the deliberate choice behind the three layers above: own the data and the ground, and treat the model as a commodity others can pour billions into and still have it switched off from outside.

For an Indigenous-language model the choice is even clearer, because the value is in the language data, not the model, and the large companies will never serve a language like Anishinaabemowin well enough to bother. That is precisely the ground a community-governed effort can hold.

Canada is already moving this way

This is not a fringe view. Canada's national AI strategy includes a sovereign-compute plan and names Indigenous self-determination over how AI is built and used. The largest Canadian model company has released open-weight models it markets for sovereign infrastructure. The argument OIATC has been making, that who governs the data and the ground matters more than who builds the model, is now close to government and industry consensus. OIATC's place in that is specific and local: the Ontario Indigenous voice for First Nations communities, and a delivery partner that already runs sovereign tools on its own infrastructure.

The dependency is real. The answer a community can actually own is its own governed data, open models it can run, and infrastructure it controls.

What a First Nation can take from this

• Treat any one company's model as replaceable. Prefer tools that let you change the model underneath without losing your data.
• Keep the community's data under OCAP, and know which systems have a copy of it.
• For anything that matters, ask what it would take to run it on infrastructure the community controls.
• Judge a vendor not only by how good the model is today, but by whether you keep your data and your access if the vendor, or a foreign government, changes the terms tomorrow.

Sources

• The company's own statement on the directive, and reporting from CNBC.
• Cohere's open-weight Command A+, built for sovereign infrastructure.
• Canada's national AI strategy, AI for All, and the Canadian Sovereign AI Compute Strategy.
• The First Nations Principles of OCAP (FNIGC).
• OIATC's companion explainer, where your community's data actually lives.

More from OIATC: where your data lives · the Anokii program · council positions